Technical improvement project text intelligent question answering method and system based on large model

Abstract

The embodiment of the invention provides a technical improvement project text intelligent question answering method and system based on a large model, and the method comprises the steps: collecting corpora of the power industry, carrying out the field adaptation of a general large model through a transfer learning or fine tuning technology, and constructing a special model for the power field; the method comprises the following steps: collecting multi-source heterogeneous data of an electric power project, carrying out data cleaning processing, key information extraction and classification integration on the multi-source heterogeneous data, and constructing a structured knowledge base by adopting a knowledge graph technology; embedding the structured knowledge base into a power field special model, and optimizing reasoning logic through reinforcement learning to obtain a power field special question and answer model; and inputting the to-be-answered question into the power field special question and answer model to obtain a corresponding professional answer. According to the technical scheme provided by the invention, the efficiency of querying professional knowledge and solving technical problems by electric power project personnel can be remarkably improved, and the professionality and accuracy of question and answer contents are guaranteed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence in the power industry, and particularly relates to a technical improvement project text intelligent question and answer method and system based on a large model. BACKGROUND

[0002] Under the background of digital transformation of the power industry, power projects involve massive technical documents, specifications, standards and engineering reports, covering planning and design, equipment operation and maintenance, fault diagnosis and other links. However, traditional manual retrieval and question and answer methods based on keyword matching cannot cope with the characteristics of complex and diverse text data, dense professional terms and scattered information in the power field, resulting in low efficiency of technical personnel in querying professional knowledge and solving technical problems, and easy to miss information or misjudge, which seriously restricts the operation and management efficiency of the power system.

[0003] At present, although general intelligent question and answer systems have been widely used, there are two technical bottlenecks due to the lack of knowledge adaptation and optimization for the power industry: first, general large models lack semantic understanding ability of professional terms in the power field, and are prone to ambiguity in analyzing problems such as "ultra-high voltage direct current transmission control strategy"; second, multi-source data (such as PDF, XML, database, etc.) in the power industry is difficult to be effectively integrated and structured, resulting in insufficient accuracy and poor timeliness of knowledge retrieval.

[0004] Therefore, it is urgent to develop an intelligent question and answer system that deeply integrates power field knowledge and large model technology to meet the urgent needs of the industry for efficient and accurate knowledge services. SUMMARY

[0005] The technical improvement project text intelligent question and answer method and system based on a large model provided by the embodiments of the application can significantly improve the efficiency of power project personnel in querying professional knowledge and solving technical problems, and guarantee the professionalism and accuracy of the question and answer content.

[0006] In a first aspect, the embodiments of the application provide a technical improvement project text intelligent question and answer method based on a large model, comprising:

[0007] Collecting power industry corpus and adapting general large models to the field through transfer learning or fine-tuning technology to build power field special models;

[0008] Collecting multi-source heterogeneous data of power projects and performing data cleaning processing on the multi-source heterogeneous data;

[0009] Extracting key information from the cleaned multi-source heterogeneous data and converting the key information into structured data;

[0010] Classifying and integrating the structured data through intelligent algorithms, and constructing a structured knowledge base using knowledge graph technology;

[0011] embedding the structured knowledge base into a power field special model, and optimizing the reasoning logic through reinforcement learning to obtain a power field special question and answer model;

[0012] inputting the problem to be solved into the power field special question and answer model to obtain a corresponding professional answer.

[0013] In an optional implementation, collecting the power industry corpus includes: collecting power grid planning and design documents, power equipment operation and maintenance manuals, new energy power generation technology reports, technical transformation project approval documents, construction schemes, acceptance reports, and fault handling records based on the whole life cycle of technical transformation projects;

[0014] The power industry corpus is screened to remove duplicate and irrelevant corpus, and the engineering units and parameter formats are unified to generate a standardized power industry corpus.

[0015] In an optional implementation, the cosine similarity algorithm is used to remove duplicate corpus, and the similarity threshold is 0.9; the expert experience method is used to remove irrelevant corpus.

[0016] In an optional implementation, the domain adaptation of the general large model through transfer learning or fine-tuning technology includes:

[0017] According to the size of the corpus, the technical path is selected:

[0018] When the corpus size is less than a first preset value, the transfer learning technology is used, the bottom general language understanding module of the general large model is retained, and the top structure is replaced with a power field special structure;

[0019] When the corpus size is greater than or equal to the first preset value, the fine-tuning technology is used, the parameters of the target layer of the general large model are frozen, and the parameters of other layers are adjusted.

[0020] In an optional implementation, the multi-source heterogeneous data is processed through data cleaning, including:

[0021] The multi-source heterogeneous data is processed through data cleaning, including:

[0022] Remove duplicate values, remove outliers, and fill in missing values.

[0023] The missing value interpolation includes:

[0024] The mean value interpolation method is used for numerical value type parameters, and the mode interpolation method is used for category type parameters.

[0025] In an optional implementation, the structured data is classified and integrated through an intelligent algorithm, specifically including:

[0026] For the text with a length less than a second preset value, a TextCNN model is used to classify according to the device model, fault type and technical improvement stage;

[0027] For the text with a length greater than or equal to the second preset value, a BERT model is used to classify according to the industry standard grade, construction safety grade and acceptance index category.

[0028] In an optional implementation, the structured knowledge base is embedded into a power field special model, specifically including:

[0029] A first vector is obtained according to the entities and relationships in the structured knowledge base;

[0030] A second vector with the same dimension as the first vector is obtained by converting the to-be-answered question and the text in the standardized power industry corpus;

[0031] The first vector and the second vector are spliced to serve as input features of the power field special model.

[0032] In a second aspect, the embodiments of the present application provide a technical improvement project text intelligent question-answering system based on a large model, including:

[0033] A data acquisition module is configured to collect power industry corpus, and adapt a general large model to a field through transfer learning or fine-tuning technology to construct a power field special model;

[0034] A data processing module is configured to collect multi-source heterogeneous data of a power project, and perform data cleaning processing on the multi-source heterogeneous data; and extract key information from the cleaned multi-source heterogeneous data, and convert the key information into structured data;

[0035] A structured knowledge base construction module is configured to classify and integrate the structured data through an intelligent algorithm, and construct a structured knowledge base by using a knowledge graph technology, and establish a dynamic updating mechanism;

[0036] A question-answering model construction module is configured to embed the structured knowledge base into a power field special model, and optimize reasoning logic through reinforcement learning to obtain a power field special question-answering model;

[0037] An output module is configured to input a to-be-answered question into the power field special question-answering model to obtain a corresponding professional answer.

[0038] In a third aspect, the embodiments of the present application provide an electronic device including a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the method provided by the embodiments of the present application.

[0039] In a fourth aspect, the embodiments of the present application provide a computer readable storage medium, characterized in that a computer program is stored on the computer readable storage medium, and when the computer program is executed in a computer, the computer is enabled to execute the method provided by the embodiments of the present application.

[0040] The technical scheme provided by the embodiments of the present application has the following beneficial effects:

[0041] The power special model is constructed by the migration learning and the field fine-tuning technology, the limitations of the traditional question and answer system in the professional term understanding are broken through, the model can deeply analyze the semantic connotation of complex power terms such as "ultra-high voltage direct current transmission control strategy", the multi-source heterogeneous power engineering data is converted into a structured semantic network by combining the intelligent classification algorithm and the knowledge graph technology, and an interconnected professional knowledge system is formed, the knowledge base and the model are deeply coupled by the double embedding vector fusion technology, so that the system can consider the text context and the professional knowledge simultaneously when processing the user's question, and then the dynamic optimization mechanism of the reinforcement learning is combined, the reasoning logic of the model is continuously trained based on the professional compliance reward function, so that the model can automatically combine the multi-modal knowledge to generate a hierarchical solution when facing complex scenes such as "typhoon weather fault disposal", and the technical means jointly act to significantly improve the response efficiency of the professional knowledge query, and optimize the traditional manual retrieval process into instant intelligent response, through the triple guarantee mechanism of knowledge matching, specification checking and feedback iteration, it is ensured that the output answer completely meets the power industry standard specification, and relying on the dynamic updating and self-evolution ability, the system can synchronize the latest industry standards and technical dynamics in real time, and continuously optimize the knowledge base and the reasoning path, the scheme provides intelligent support with professional depth and engineering feasibility for the whole life cycle of the power technical improvement project from the project establishment, construction to acceptance, and effectively promotes the quality and efficiency of the digital transformation of the power industry. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a process schematic diagram of the text intelligent question and answer method for technical improvement projects based on a large model provided by the embodiments of the present application;

[0043] Figure 2 is a structure schematic diagram of the text intelligent question and answer system for technical improvement projects based on a large model provided by the embodiments of the present application. DETAILED DESCRIPTION

[0044] The present application is further described in detail below by means of the drawings and the specific embodiments.

[0045] The embodiments of the present application provide a text intelligent question and answer method for technical improvement projects based on a large model, Figure 1FIG. 1 is a flowchart of a method for intelligent question answering of technical transformation project text based on a large model according to an embodiment of the present application. The method can be executed by a system for intelligent question answering of technical transformation project text based on a large model. The system can be implemented by software and / or hardware, and can be configured in a computer or other electronic device.

[0046] As shown in FIG. 1, the technical solution provided by the present application includes the following steps: Figure 1

[0047] S110, collect power industry corpus, and adapt a general large model to a power field through migration learning or fine-tuning technology to build a power field special model.

[0048] Specifically, collecting power industry corpus includes: collecting power grid planning and design documents, power equipment operation and maintenance manuals, new energy power generation technology reports, technical transformation project approval documents, construction schemes, acceptance reports and fault handling records based on the whole life cycle of technical transformation projects; filtering the power industry corpus to remove duplicate and irrelevant corpus, and unifying engineering units and parameter formats to generate a standardized power industry corpus library. The power industry corpus library refers to all text materials carrying professional knowledge in the field of power transformation, which is the basic data source for building a structured knowledge base, covering technical specification documents (such as DL / T573-2010 "Power Transformer Maintenance Guidelines") in the power industry, technical transformation project life cycle documents (approval documents, construction schemes, acceptance reports), equipment operation and maintenance manuals (such as "220kV Main Transformer Maintenance Process"), fault handling records (such as "GIS Device SF6 Leakage Disposal Log"), professional term explanations (such as "Differential Protection 87L Definition"), etc.

[0049] First, a wide range of corpus in the power industry needs to be collected to ensure the richness and professionalism of the corpus. The corpus collection range covers the whole life cycle of technical transformation projects, including but not limited to:

[0050] Power grid planning and design documents: such as power grid upgrade scheme, transmission line layout design, etc.

[0051] Power equipment operation and maintenance manuals: such as maintenance and repair guidelines for transformers, circuit breakers, relay protection devices, etc.

[0052] New energy power generation technology reports: such as technical analysis and operation data of wind and solar power generation projects;

[0053] Technical transformation project approval documents, construction schemes, acceptance reports and fault handling records: these documents are directly related to the implementation process of technical transformation projects, and contain a large number of professional parameters and specifications.

[0054] In order to ensure the quality of the corpus, a three-level audit mechanism is used for quality control:

[0055] ​First-level review: Manually screened by power industry experts, eliminating general power documents unrelated to technical transformation, ensuring that the corpus focuses on the technical transformation field.

[0056] Second-level review: Use cosine similarity algorithm for deduplication, set the similarity threshold to 0.9. The threshold selection is based on the professional characteristics of the power technical transformation corpus: if the threshold is too low (such as 0.8), it may cause the deletion of similar expressions but different core information (such as technical transformation parameters of different types of transformers), missing key technical details; if the threshold is too high (such as 0.95), it will retain too many duplicate corpora (such as different copies of the same specification), causing redundancy. Threshold 0.9 can balance the deduplication effect and information integrity, only deleting highly repetitive corpora (such as completely identical equipment operation clauses), and retaining effective corpora with slightly different expressions but different core information.

[0057] Third-level review: Manually sampled and checked at a rate of 5% to ensure that professional terms (such as "differential protection 87L" and "harmonic distortion rate THD") are accurately and unambiguously expressed.

[0058] For non-digital text (such as handwritten operation records, paper technical transformation logs), use OCR (Optical Character Recognition) technology for digital extraction. OCR technology uses optical pattern recognition algorithms to capture character shapes and arrangement patterns in images, converting text (including Chinese characters, numbers, and professional symbols such as "kV" and "℃") into editable digital text while removing noise (such as paper crease marks). Subsequently, use NLTK (Natural Language Processing Toolkit) to process the text, breaking it down into the smallest language units (such as words or phrases) based on semantic logic. In the power context, NLTK will prioritize recognizing professional terms (such as "220kV transformer"), industry standard numbers (such as "DL / T573-2010"), and technical parameters, avoiding errors caused by general word segmentation tools (such as splitting "DL / T573-2010" into meaningless fragments). The segmented corpus is then standardized using regular expression templates to engineer units and parameter formats, resulting in a standardized training corpus.

[0059] Further, according to the corpus size and task complexity, choose transfer learning or fine-tuning technology to adapt general large models (such as GPT-3.5, LLaMA) to the power field. The selection criteria are as follows:

[0060] Transfer learning: Suitable for smaller corpus size (such as less than 250,000) and general large models without pre-training in the power field. Its core is to reuse the underlying general language understanding capabilities of general large models, only adjusting the upper structure to adapt to the power task.

[0061] Specific implementation:

[0062] Retain the bottom general-purpose module: In the 12-layer Transformer encoder of the general large model (such as BERT-base), the bottom layers (the first 4-6 layers) are responsible for general language basic semantic understanding (such as word-level feature recognition, local syntax structure parsing), and these capabilities do not depend on specific domain knowledge. Therefore, the core parameters of the bottom module are retained to ensure that the model can first "understand" the general language (such as the general expression in "help me explain this construction plan").

[0063] Modify the upper network: Replace the top classifier with a power-specific structure based on IEC standards, such as modifying the output layer to a classification head that can identify "power failure types" and "technological improvement step compliance". At the same time, use local maximum mean difference (LMMD) constraints for sub-domain adaptation, and set the feature norm constraint value to Δr=0.5 to align the bottom output features with the feature distribution of power domain data (such as power generation, transmission, and distribution sub-domains), avoiding semantic disconnection.

[0064] Fine-tuning technology: suitable for cases with sufficient corpus size (such as more than 250,000) and complex tasks (such as fault root cause analysis). The core is to adjust all or part of the model parameters to adapt to the power scene.

[0065] Specific implementation:

[0066] Parameter adjustment strategy: freeze the first 8 layers of Transformer parameters (retain general-purpose capabilities), and adjust all parameters from the middle to the top layer. For specific tasks (such as power failure diagnosis and question answering, engineering budget consultation), select labeled training datasets (such as "technological improvement fault type" and "device modification priority" labels) to optimize the model.

[0067] Training settings: set the initial learning rate to 2e-5, the batch size to 64, and perform 2-3 training cycles on a GPU cluster.

[0068] In the process of training the large model, a hybrid loss function and an attention regularization loss function are used to optimize the model performance:

[0069] Hybrid loss function: combine negative log likelihood loss (NLLLoss) and center loss to enhance the model's discriminative ability for classification tasks.

[0070] Attention regularization loss function, formula:

[0071] ;

[0072] Where M represents the total number of local features, f k is the kth local feature extracted by the model (such as the "transformer oil temperature" parameter identified from the text), and c kis the target center corresponding to the k-th local feature, which is pre-set by power experts (e.g., the standard value of "oil level indication of oil-immersed transformer"). 2 represents the square of the Euclidean distance, which is used to measure the deviation of the feature from the target center. This function forces the model to focus on the key information of power (such as equipment parameters, technical specifications) by minimizing the deviation, avoiding dispersion to irrelevant content.

[0073] After training, the model can be verified in the following ways:

[0074] Cross-validation: Ensure that the Top-1 accuracy of the model in the power professional task is ≥ 96.875%, and the performance fluctuation of the sub-field (such as substation operation and maintenance, distribution automation) is ≤ 3%.

[0075] Feature visualization: Use t-SNE dimensionality reduction technology to visualize the feature space and verify the clustering effect of "equipment abnormality description" and "standard solution".

[0076] Attention heat map analysis: Confirm that the model can accurately focus on key equipment parameters (such as the number of circuit breaker opening and closing).

[0077] Finally, the model is packaged as a RESTful API service that meets the IEC61970 standard, supporting input of unstructured power work order text and output of standardized results that meet the CIM / E specification.

[0078] S120, collect multi-source heterogeneous data of power projects, and perform data cleaning processing on the multi-source heterogeneous data.

[0079] Specifically, the data of power technical improvement projects comes from various sources and has different formats. First, different types of data need to be extracted and preliminarily organized using appropriate tools. For example, the following categories can be included:

[0080] PDF document processing: For technical specifications, acceptance reports, and other PDF format files, use Apache Tika tool library to extract text content, and combine PDFBox tool to remove watermarks, headers, footers, and other redundant information, and keep the core text.

[0081] XML file processing: For device parameter configuration, project progress data, and other XML files, use DOM4J parser to read its hierarchical data structure, and then convert it to flat table data through XSLT (Extensible Stylesheet Language Transformation) technology for subsequent processing.

[0082] Database data integration: For the device operation records, fault statistics and other data stored in databases such as MySQL, Oracle, etc., use SQL query statements to filter valid data fields, and use ETL (Extraction, Transformation, Loading) tools such as Kettle for preliminary integration, eliminate redundant fields such as device ID, and unify date fields to "YYYY-MM-DD" standard format.

[0083] Unstructured text processing: For on-site handwritten operation records, technical improvement scheme drafts and other non-digital or disorganized text, after being converted into digital text through OCR technology, use NLTK (Natural Language Toolkit) for word segmentation, and use regular expression templates to extract core information such as "fault time" and "treatment measures", and filter out irrelevant chitchat or notes.

[0084] Further, after preliminary data extraction, enter the core data cleaning phase to improve data quality:

[0085] Duplicate data identification and deletion: Use Jaccard similarity coefficient algorithm to calculate the similarity between text records. The coefficient is defined as the ratio of the intersection size of two sets to the union size, formula:

[0086] J(A,B)=(|A∩B|) / (|A∪B|);

[0087] Where J(A,B) represents the Jaccard similarity coefficient of set A and set B, A and B represent the word sets formed after word segmentation of two data records.

[0088] Exemplarily, set the threshold value to 0.85, when the Jaccard coefficient of two records ≥0.85, it is determined as highly repetitive record, the system will automatically delete the duplicate. This threshold value can effectively eliminate records with almost the same content (such as repeatedly uploaded same device inspection report), while retaining records with different core information (such as device parameters, fault codes) although some descriptions are the same.

[0089] Error data detection and elimination: For numerical parameters (such as device temperature, voltage value), use Z-score (standard score) algorithm for outlier detection. Z-score measures the deviation of a data point from the mean of the data set, formula:

[0090] z=(x-μ) / σ;

[0091] Where z represents Z-score (standard score), x is the original data point, μ is the mean of the data set, and σ is the standard deviation of the data set.

[0092] Exemplarily, the absolute value of the Z-score is set as 3 (i.e. ±3) as the threshold. When |z|>3 for a certain data point, it is determined as an error data significantly deviating from the normal range (for example, the normal transformer oil temperature is about 65-85℃, and if a record of 150℃ appears, it is regarded as abnormal), and is rejected.

[0093] Missing value processing: different imputation strategies are adopted according to the data type of the missing field:

[0094] Mean imputation: suitable for missing of numerical fields (such as device load rate, construction duration). The arithmetic mean value of all valid data of the field is calculated, and the value is used to fill in the missing value. For example, the load rate data of a certain transformer in other months is 65%, 72%, 68%, and 70%, and the mean value is (65+72+68+70) / 4=68.75%, which is used to fill in the missing month load rate. This method can maintain the overall statistical characteristics of the data.

[0095] Mode imputation: suitable for missing of categorical fields (such as fault type, device model, acceptance conclusion). The category with the highest frequency (i.e. mode) among all valid data of the field is counted, and the category is used to fill in the missing value. For example, in the acceptance record of technical transformation project, the “qualified” appears 85 times, “to be rectified” appears 12 times, and “unqualified” appears 3 times in the “acceptance level” field, and the mode is “qualified”, which is used to fill in the missing value. This method conforms to the actual distribution of category data.

[0096] Special processing: if the missing value proportion of a certain field is as high as 30% or more, it is considered that the data quality of the field is too low, and the whole record (sample) is directly deleted to avoid introducing too large deviation.

[0097] S130, key information extraction is performed on the cleaned multi-source heterogeneous data, and the key information is converted into structured data.

[0098] Specifically, key information extraction mainly relies on information extraction technology in natural language processing (NLP) to identify entities with specific meanings from text and clarify the relationships between them.

[0099] Named entity recognition (NER): specific named entities are identified and classified from text. In the field of power technical transformation, these entities mainly include:

[0100] Power equipment name: such as “220kV main transformer” and “SF6 circuit breaker”.

[0101] Technical parameters: such as “rated capacity 2500kVA” and “insulation withstand voltage ≥30kV”.

[0102] Operation process: such as “insulation resistance test” and “protection setting value verification”.

[0103] Safety standards: as per “DL / T 573-2010 Guide for Overhaul of Power Transformer” and “GB 50150-2016 Standard for Electrical Equipment Interim Test in Electrical Installation Engineering”.

[0104] NER (Named Entity Recognition) is performed using the BERT-CRF model. The BERT (Bidirectional Encoder Representations from Transformers) model is responsible for deep understanding of the context semantics of the text, generating context-related vector representations for each character. The CRF (Conditional Random Field) model serves as the output layer, utilizing sequence labeling (such as BIO, i.e., Begin, Inside, Outside) strategies, and considering label constraints at the sentence level, to ensure the rationality and global optimality of the output entity label (such as “equipment”, “parameter”, “standard”) sequence.

[0105] For terms like “grounding” that have different meanings in different contexts, the model will disambiguate and accurately label them based on their context semantics. For example, in “the protective grounding resistance value should be qualified”, it is labeled as “grounding-protection measures”; in “a single-phase grounding fault occurs”, it is labeled as “grounding-fault type”.

[0106] Relation Extraction: Based on the identified entities, further extract the semantic relationships between entities to build a network of entity relationships.

[0107] Adaptive instruction decoding technology such as REAREV (Relational Extraction with Adaptive Relation Vectors) is used. This technology learns specific relationship vectors between entity pairs by predefining relationship types (such as “must meet”, “reference”, “belongs to”), thereby establishing associations between “equipment-parameters-standards” from the text.

[0108] Exemplarily, from the sentence “The rated capacity of the 220kV main transformer is 2500kVA, which must meet the DL / T 573-2010 standard”, the NER model identifies the entities “220kV main transformer” (equipment), “rated capacity 2500kVA” (parameter), and “DL / T 573-2010” (standard). The relation extraction model further extracts two sets of relationships: (220kV main transformer, has parameter, rated capacity 2500kVA) and (220kV main transformer, must meet standard, DL / T 573-2010).

[0109] Further, the extracted key information needs to be structured according to a unified data format and stored in a database or knowledge graph to form a standardized data storage system.

[0110] Structured processing and storage:

[0111] Relational database storage: Store entities and their attributes in table form. For example, create a "Technical Improvement Project Core Information Table", and the core fields can be designed as follows:

[0112] Project ID (primary key, such as JGY2025001), project name (such as "XX Substation 220kV Main Transformer Technical Improvement Project"), equipment ID involved (foreign key, associated with "Equipment Information Table"), reference standard ID (foreign key, associated with "Technical Specification Table"), key parameters (such as "Rated capacity: 2500kVA"), etc.

[0113] Knowledge graph (RDF format) storage: To more clearly express the complex association between entities, use Resource Description Framework (RDF) format to store knowledge. RDF uses a "subject-predicate-object" triple form to express knowledge, which is very suitable for representing entity relationships extracted from text. Store the triple of entities and relationships in RDF format, or use the attribute graph model to store the attributes and associations of knowledge nodes, and store them in a special graph database (such as Neo4j). The core function of this storage method is to enable the subsequent system to quickly "find" knowledge: when the user asks "220kV transformer technical improvement insulation oil test standard", the system can quickly locate the association between "220kV transformer" and "insulation oil test" "DL / T573-2010" through the stored graph structure, without traversing massive original texts, greatly improving the knowledge matching efficiency.

[0114] Example: Convert the results of the above relationship extraction into RDF triples:

[0115] Subject: 220kV main transformer, predicate: has parameter, object: rated capacity 2500kVA.

[0116] Subject: 220kV main transformer, predicate: needs to meet standards, object: DL / T573-2010.

[0117] This storage method forms a semantic network, which is convenient for deep graph retrieval and reasoning.

[0118] S140, classify and integrate structured data through intelligent algorithms, and build a structured knowledge base using knowledge graph technology.

[0119] Specifically, first, for the key information obtained in S130 (such as device entities, parameters, and standards), it is necessary to classify them in more detail according to their attributes and application scenarios for easy management and retrieval.

[0120] Automatic classification:

[0121] Reference threshold for short texts: in terms of the number of characters, the number of characters in a single text is ≤500; in terms of the number of words after power professional field segmentation, the number of words is ≤100. Such texts correspond to device parameter descriptions in technical improvement scenarios (such as "SF6 circuit breaker gas pressure normal value 0.6-0.7 MPa, low pressure alarm value 0.55 MPa"), fault code explanations (such as "Fault code E01: transformer oil temperature sensor is broken, the processing method is to replace the sensor and calibrate"), and brief descriptions of technical improvement stages (such as "During the project initiation stage, a project feasibility analysis report needs to be submitted to the operation and maintenance department for review"). The core feature of these texts is focused information and single semantics, with no complex cross-paragraph logical associations.

[0122] Classification of short texts: For device parameter descriptions and fault code explanations, which are relatively short in length, the TextCNN (Text Convolutional Neural Network) model is used for classification. TextCNN uses different size convolution kernels to extract local features of the text (such as n-gram features), which can effectively capture key classification dimensions such as "device model", "fault type", and "technical improvement stage". Through training, the classification accuracy can reach 92%-94%.

[0123] Reference threshold for long texts: in terms of the number of characters, the number of characters in a single text is >500; in terms of the number of words after power professional field segmentation, the number of words is >100. Such texts correspond to full texts of power technical improvement technical specifications (such as DL / T573-2010 "Guidelines for Power Transformer Maintenance"), technical improvement project acceptance reports (including construction process, detection data, acceptance conclusion, and other modules), and chapters of device operation and maintenance manuals (such as "220kV main transformer maintenance process and quality standards"). The core feature of these texts is multiple information dimensions, containing cross-sentence logical dependencies (such as the association between construction steps and the corresponding relationship between detection data and acceptance standards), and higher complexity of context association of professional terms.

[0124] Classification of long texts: For long texts such as technical specifications and project acceptance reports, the BERT (Bidirectional Encoder Representations from Transformers) model is used for fine-grained classification. BERT, with its powerful bidirectional context understanding ability, can more accurately classify according to "industry standard levels" (such as international standards, industry standards, and enterprise standards), "construction safety levels" (such as first-level risk and second-level risk), and "acceptance index categories" (such as performance indicators and safety indicators), with an accuracy rate of 95%-97%.

[0125] In some embodiments, the above classification threshold is not an absolute rigid standard and needs to be adjusted in combination with the professional information density of the power technical improvement text. For example, a text with 600 characters may only contain highly focused professional expressions such as "action conditions, setting range, and fault handling process of differential protection 87L" with no redundant information, and its semantic complexity is similar to that of short texts. The system will classify it as a short text and use the TextCNN model to process it. Conversely, if a text has 450 characters but contains multiple dimensions of information such as "safety measures, personnel configuration, material list, and project planning" and has complex logical associations, the system will classify it as a long text and use the BERT model to process it. This dual classification method of "numerical threshold + professional semantic complexity" is designed to adapt to the technical characteristics of TextCNN and BERT models - TextCNN is good at capturing local features of short texts and has high efficiency and accuracy in short text classification, with an accuracy rate of 92%-94%. BERT relies on a bidirectional attention mechanism to handle long-distance semantic dependencies in long texts, with an accuracy rate of 95%-97% in fine-grained long text classification, ultimately ensuring the accuracy and efficiency of multi-source data classification to meet system requirements.

[0126] Classification result calibration: All classification results are calibrated through 5-fold cross-validation, i.e., dividing the data set into 5 parts, using 4 parts for training and 1 part for testing, repeating 5 times to take the average accuracy rate, to effectively avoid overfitting or underfitting problems caused by improper data division, ensuring the stability and reliability of classification, and preventing misclassification and missed classification.

[0127] Further, the classified data needs to be integrated through knowledge graph technology to connect isolated entities and relationships into a large semantic network, forming a real structured knowledge base.

[0128] Construction of knowledge graph:

[0129] The knowledge graph takes "technical transformation project entities" (such as project number, equipment name, construction unit) and "technical specification entities" (such as standard number, parameter requirement, acceptance index) as core nodes. Then, based on the results of relationship extraction in S130, the association between nodes is established.

[0130] Example: Taking "220kV transformer technical transformation project" as the central node, the following can be associated:

[0131] Standard node: "DL / T573-2010 Power Transformer Maintenance Guidelines".

[0132] Equipment node: "S11-2500kVA transformer".

[0133] Parameter node: "Insulating oil withstand voltage test ≥ 30kV".

[0134] The established association includes: "Project A-Reference-Standard B", "Equipment C-Need to meet-Parameter D", "Construction step E-Dependence-Condition F", etc. These relationships clearly show the logical association between knowledge nodes in the knowledge base, forming a visual knowledge network.

[0135] Knowledge representation and storage: The knowledge graph is usually stored in RDF (Resource Description Framework) format, that is, in the form of "subject-predicate-object" triplets. For example:

[0136] Subject: 220kV transformer technical transformation project, predicate: involves equipment, object: S11-2500kVA transformer.

[0137] Subject: S11-2500kVA transformer, predicate: need to meet parameters, object: insulating oil withstand voltage test ≥ 30kV.

[0138] Subject: Insulating oil withstand voltage test ≥ 30kV, predicate: reference standard, object: DL / T573-2010.

[0139] This storage method not only has high degree of structuring, but also is more convenient for graph traversal and complex reasoning queries (such as multi-hop queries).

[0140] In addition, in order to ensure the timeliness of the knowledge base and keep up with the rapid development of the power industry, the embodiment also establishes a dynamic updating mechanism.

[0141] Update trigger conditions and processes:

[0142] Regular update: Regular equipment maintenance data is updated once a month; industry specifications and technical standards are checked once a quarter.

[0143] Active monitoring and triggered updates: The system monitors power industry dynamics in real-time, including:

[0144] Changes in policies and regulations: such as the release of new DL / T standards.

[0145] Development of new technologies: such as the application of new technologies such as flexible DC devices.

[0146] Internal system indicators: when the knowledge base query error rate is ≥5% continuously, or when users feedback that key information is missing.

[0147] Once the above conditions are monitored, the update process is triggered immediately. New materials are completed within 72 hours of review, structured processing (going through the process of S120-S130), and integration with the knowledge base.

[0148] Version management: version management of the knowledge base to ensure traceability of updates and, if necessary, rollback to previous stable versions.

[0149] S150, embed the structured knowledge base into the power field special model, and optimize the inference logic through reinforcement learning to get the power field special question and answer model.

[0150] Specifically, in order to enable the large model to directly call and follow professional knowledge when processing user questions, it is necessary to deeply embed the structured knowledge base constructed in S140 into the power field special model trained in S110.

[0151] This embodiment adopts a double embedding fusion scheme, not simply concatenating the three vectors, but implementing the fusion of two types of vectors "knowledge graph vector + text vector" based on the double embedding scheme:

[0152] The first vector (knowledge graph feature vector): the "entity-relation" vector of the knowledge graph (not a single vector). Use knowledge graph embedding algorithm (such as TransE) to convert the "entity-relation" triplets in the knowledge base into low-dimensional, dense vector representations. For example, the triplet (device A, needs to meet, parameter B) is mapped to a set of 512-dimensional vectors (v A , v 关系 , v B ). This process allows symbolic knowledge to be understood and processed by mathematical models. Entity A, relationship, and entity B form a feature vector group for the triplet, representing structured knowledge associations. If multiple entity-relation triplets are involved (such as "technological improvement project C-referenced-standard D" "technological improvement project C-involved-device A"), the vector groups of multiple triplets are fused to generate a unified knowledge graph feature vector.

[0153] Second vector (text feature vector): The text vector of the question to be answered and the text in the standardized power industry corpus (two independent vectors are fused). The user's natural language question (such as "How to handle the 220kV main transformer oil temperature overrun?") and the text segment in the knowledge base (such as the content of the standard clause) are also converted into semantic vectors of the same dimension (512 dimensions) using the Sentence-BERT model. Then the question and the text are fused (such as splicing, addition or attention weighting) to generate a single "text feature vector".

[0154] Feature splicing and input: The fused knowledge graph feature vector and the fused text feature vector are spliced to form the input feature vector of the large model, rather than directly splicing the "question vector, standardized power industry corpus text vector, entity-relation vector" three independent vectors. The core of this splicing method is to enable the model to capture both "structured entity-relation association" and "natural language text context information", ensuring that the answer is consistent with the professional knowledge association of power technical transformation, and understanding the semantic intent of the user's question, and finally generating accurate professional question and answer results.

[0155] Further, on the basis of integrating knowledge, the "decision" ability of the model is trained through reinforcement learning, so that it learns how to filter and combine information from a large amount of knowledge to generate answers that conform to the logic of power engineering.

[0156] The core of reinforcement learning is to design a scientific reward (Reward) mechanism to evaluate the good and bad of the model's generated answers and guide its optimization. For example, the reward function R can be constructed based on the following rules:

[0157] Positive reward:

[0158] Answer professionalism (+2 points): If the generated answer accurately quotes key information in the knowledge base, such as standard clauses ("in accordance with DL / T 573-2010 Article 5.2") and device parameters ("rated capacity 2500kVA").

[0159] Logical compliance (+1.5 points): If the reasoning steps of the answer conform to the power engineering process, such as "first detect insulation resistance, then test voltage strength".

[0160] User satisfaction (+3 points): If the user gives positive feedback that the answer is "accurate and useful".

[0161] Negative reward (penalty):

[0162] Knowledge conflict (-3 points): If the information in the answer conflicts with the knowledge base record, such as incorrect parameters.

[0163] Logical violation (-5 points): If the reasoning logic seriously violates the industry safety specifications, such as the suggestion "disassemble the device without power off".

[0164] User denial (-4 points): If the user feedback answer is "wrong" or "irrelevant".

[0165] Optimization algorithm and process:

[0166] Advanced reinforcement learning algorithms such as PPO (Proximal Policy Optimization) are used. This algorithm continuously tries to generate answers, calculates rewards, and updates model parameters to make the model strategy (i.e. answer generation method) evolve towards obtaining higher cumulative rewards. The ultimate goal is to enable the model to automatically select the most relevant knowledge nodes (such as "windproof plan", "fault isolation procedure") and combine them into a logically rigorous and highly feasible hierarchical solution when faced with complex multi-factor problems such as "distribution network fault handling priority in typhoon weather".

[0167] In addition, to ensure the continuous evolution of the system, this embodiment also establishes a dynamic adjustment mechanism for closed-loop freshness.

[0168] Key indicator monitoring: The system records real-time "user question - model answer - called knowledge - user feedback" full-link data. Two key thresholds are set:

[0169] Error feedback rate threshold (8%): Triggered when the negative feedback rate of a certain type of problem (such as "GIS device technical improvement acceptance") is ≥8% for 3 consecutive days.

[0170] Knowledge call missing rate threshold (10%): Triggered when the proportion of answers generated directly by the model without calling relevant knowledge is ≥10%.

[0171] Targeted optimization: Once triggered, immediately start the adjustment process. For example, for the "GIS device acceptance" problem, supplement relevant acceptance case corpus and fine-tune the parameters in the model layer; at the same time, optimize the knowledge base retrieval strategy and expand the retrieval range of related nodes.

[0172] Iterative verification: After adjustment, use a special test set to verify the effect. If the error feedback rate is reduced to less than 3%, the optimization is complete; otherwise, repeat the monitoring and adjustment process to ensure that the model always maintains high levels of precision matching and answer generation capabilities.

[0173] S160, input the problem to be solved into the power field special question and answer model, and get the corresponding professional answer.

[0174] For example, the user inputs the problem to be solved through the interactive interface provided by the system, and the system preprocesses the input to ensure that the model can correctly understand:

[0175] Among them, the power field special question and answer model supports multi-modal input, including:

[0176] Text input: users directly input text questions, such as "How to handle 220kV main transformer oil temperature overrun?";

[0177] Voice input: in noisy environments such as substations, users can ask questions through voice, and the system integrates noise reduction recognition technology to improve recognition accuracy;

[0178] Picture assisted input: users can upload device failure photos, and the system automatically identifies device types and associates relevant question templates.

[0179] The text input module is equipped with intelligent completion function of power professional terms. When the user inputs "transform", the system automatically prompts "transformer technical improvement", "transformer parameters" and other options; The input content is standardized to ensure consistency with the terminology in the knowledge base.

[0180] The power field special question and answer model deeply understands the input question and retrieves relevant information in the knowledge base:

[0181] Semantic analysis and intent recognition:

[0182] The model first analyzes the semantic structure of the question and identifies the core intent. For example, for the question "How to handle 220kV main transformer oil temperature overrun?", the model identifies:

[0183] Device type: 220kV main transformer;

[0184] Question type: oil temperature overrun;

[0185] Demand type: query for treatment scheme.

[0186] Knowledge base collaborative retrieval:

[0187] Dual retrieval mechanism is adopted:

[0188] Vector retrieval: convert the question into a semantic vector and perform similarity matching in the vector space of the knowledge base;

[0189] Graph structure retrieval: traverse the path of "device-failure-solution" in the knowledge graph for query;

[0190] The retrieval range includes: relevant technical standards, similar fault cases, device parameter requirements, safety operation procedures, etc.

[0191] The model generates professional answers based on the retrieved knowledge through multi-step reasoning:

[0192] Multi-step reasoning mechanism:

[0193] Knowledge filtering: filtering out the most relevant knowledge pieces from the search results for the current question;

[0194] Logical combination: combining knowledge pieces according to the standard process of power engineering. For example, for the problem of oil temperature exceeding the limit, the answer is organized according to the logic of "cause analysis → detection steps → treatment measures";

[0195] Compliance check: ensuring that the answer content meets the requirements of DL / T, GB / T and other power industry specifications.

[0196] Quality optimization before generating the final answer:

[0197] Professional verification:

[0198] Check the correctness of technical parameters in the answer (such as voltage level, temperature range, etc.);

[0199] Verify the logical reasonableness of the solution (whether it meets the practice of power engineering);

[0200] Readability optimization:

[0201] Convert professional terms into language that users can easily understand;

[0202] Add necessary explanatory content to help users understand technical details;

[0203] Multi-format output:

[0204] Support text, voice, chart, and other forms of answer output;

[0205] Provide links or references to relevant technical standards;

[0206] Finally generate a structured professional answer.

[0207] After the answer is output, the system synchronously starts the feedback mechanism:

[0208] Instant evaluation function: users can give star ratings (1-5 stars) and label choices ("terminology error", "incomplete information", etc.) to the answer;

[0209] Feedback data analysis: the system records feedback information for subsequent model optimization and knowledge base update;

[0210] Closed-loop optimization: negative feedback will trigger the dynamic adjustment mechanism in S150 to achieve continuous improvement of the system.

[0211] Figure 2 is a structural diagram of the technical improvement project text intelligent question and answer system based on a large model provided by the embodiments of the present application, as shown in Figure 2 The system includes:

[0212] The data acquisition module is configured to collect a power industry corpus, and perform field adaptation on a general large model by using a transfer learning or fine-tuning technology, so as to construct a power field special model.

[0213] The data processing module is configured to collect multi-source heterogeneous data of a power project, and perform data cleaning processing on the multi-source heterogeneous data, extract key information from the cleaned multi-source heterogeneous data, and convert the key information into structured data.

[0214] The structured knowledge base construction module is configured to classify and integrate the structured data by using an intelligent algorithm, construct a structured knowledge base by using a knowledge graph technology, and establish a dynamic updating mechanism.

[0215] The question and answer model construction module is configured to embed the structured knowledge base into the power field special model, optimize reasoning logic by using reinforcement learning, and obtain a power field special question and answer model.

[0216] The output module is configured to input a to-be-answered question into the power field special question and answer model, and obtain a corresponding professional answer.

[0217] The execution process of the system part of the embodiment of the present application is consistent with the embodiments of the above method part, and will not be described here.

[0218] The embodiment of the present application also provides an electronic device, which comprises at least one processor and a memory connected with the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform any one of the new energy based power grid multi-resource coordination control methods.

[0219] The embodiment of the present application also provides a computer storage medium storing a computer program, and the computer program is executed by a processor to implement any one of the new energy based power grid multi-resource coordination control methods.

[0220] The computer storage medium can be referred to as a medium simply. Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The program can be stored in a non-volatile computer readable storage medium, and when the program is executed, the processes of the above-mentioned embodiments can be included. Any reference to memory, storage, database or other medium used in each embodiment provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM). Each embodiment in the specification is described in a progressive manner, and the same or similar parts of each embodiment can be referred to each other. Each embodiment focuses on the difference from other embodiments. Especially, for the device, equipment, non-volatile computer storage medium embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant part can be referred to the part of the method embodiment.

[0221] The above embodiments are only examples for clearly illustrating, and not limiting the embodiments. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the embodiments cannot be exhausted, and the obvious changes or variations still fall within the protection scope of the present application.

Claims

1. A text-based intelligent question-answering method for technical upgrading projects based on a large model, characterized in that: include: Collect corpora from the power industry and adapt general large models to specific power industry domains through transfer learning or fine-tuning techniques; Collect multi-source heterogeneous data from power projects and perform data cleaning processing on the multi-source heterogeneous data; Key information is extracted from the cleaned multi-source heterogeneous data, and the key information is transformed into structured data; The structured data is classified and integrated using intelligent algorithms, and a structured knowledge base is constructed using knowledge graph technology. The structured knowledge base is embedded into the power industry-specific model, and the reasoning logic is optimized through reinforcement learning to obtain a power industry-specific question-answering model. Input the question to be answered into the power industry-specific question-and-answer model to obtain the corresponding professional answer.

2. The method according to claim 1, characterized in that, The collected power industry corpus includes: power grid planning and design documents, power equipment operation and maintenance manuals, new energy power generation technology reports, technical upgrade project proposals, construction plans, acceptance reports, and fault handling records collected based on the entire life cycle of technical upgrade projects; The power industry corpus is screened to remove duplicate and irrelevant data, and the engineering units and parameter formats are standardized to generate a standardized power industry corpus.

3. The method according to claim 2, characterized in that, Cosine similarity algorithm was used to remove duplicate data with a similarity threshold of 0.9; expert experience method was used to remove irrelevant data.

4. The method according to claim 1, characterized in that, The domain adaptation of general large models through transfer learning or fine-tuning techniques includes: Choose the technical approach based on the size of the corpus: When the corpus size is smaller than the first preset value, transfer learning technology is used to retain the underlying general language understanding module of the general large model and replace the top-level structure with a structure specific to the power field. When the corpus size is greater than or equal to the first preset value, fine-tuning techniques are used to freeze the parameters of the target layer of the general large model and adjust the parameters of other layers.

5. The method according to claim 1, characterized in that, Data cleaning processing of the multi-source heterogeneous data includes: Data cleaning processing of the multi-source heterogeneous data includes: Remove duplicate values, remove outliers, and impute missing values; The missing value imputation includes: For numerical parameters, mean interpolation is used; for categorical parameters, mode interpolation is used.

6. The method according to claim 1, characterized in that, The classification and integration of the structured data using intelligent algorithms specifically includes: For texts shorter than the second preset value, the TextCNN model is used to classify them according to equipment model, fault type, and technical upgrade stage. For text that is greater than or equal to the second preset value, the BERT model is used to classify it according to industry standard level, construction safety level and acceptance index category.

7. The method according to claim 2, characterized in that, The embedding of the structured knowledge base into the power industry-specific model specifically includes: The first vector is obtained based on the entities and relationships in the structured knowledge base; The question to be answered and the text in the standardized power industry corpus are transformed into a second vector with the same dimension as the first vector. The first vector and the second vector are concatenated and used as input features for the power sector-specific model.

8. A text-based intelligent question-answering system for technical upgrading projects based on a large model, characterized in that: include: The data acquisition module is used to collect corpora from the power industry and adapt general large models to the power industry through transfer learning or fine-tuning techniques to build power industry-specific models. The data processing module is used to collect multi-source heterogeneous data from power projects and perform data cleaning processing on the multi-source heterogeneous data. And extract key information from the cleaned multi-source heterogeneous data, and transform the key information into structured data; The structured knowledge base construction module is used to classify and integrate the structured data through intelligent algorithms and to construct a structured knowledge base using knowledge graph technology. The question-answering model construction module is used to embed the structured knowledge base into the power field-specific model and optimize the reasoning logic through reinforcement learning to obtain a power field-specific question-answering model. The output module is used to input the question to be answered into the power industry-specific question-and-answer model and obtain the corresponding professional answer.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program, implements the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed in a computer, causes the computer to perform the method described in any one of claims 1-7.

Images